Christopher Berry
@cjpberry.bsky.social
Perpetually mystified | Views are mine.
Reposted by Christopher Berry
Great perspective from #CDSM2025: causal inference is 'what-if' analysis. You don’t have to get everything perfect—or how dare you use the c-word. What matters is laying out your assumptions transparently and showing us what happens when they're violated.
November 15, 2025 at 6:47 PM
Great perspective from #CDSM2025: causal inference is 'what-if' analysis. You don’t have to get everything perfect—or how dare you use the c-word. What matters is laying out your assumptions transparently and showing us what happens when they're violated.
My clothes are still soaked from Milton Mayer’s “They Thought They Were Free”. That bit where the Americans believed they were causing the Germans to accept responsibility for the Nazis at Nuremberg, and that 9 in 10 Nazis were completely devoid of shame.
Sticky. Sticky stuff.
Sticky. Sticky stuff.
November 15, 2025 at 4:26 PM
My clothes are still soaked from Milton Mayer’s “They Thought They Were Free”. That bit where the Americans believed they were causing the Germans to accept responsibility for the Nazis at Nuremberg, and that 9 in 10 Nazis were completely devoid of shame.
Sticky. Sticky stuff.
Sticky. Sticky stuff.
Reposted by Christopher Berry
Over on the Gelman blog, there's a post about this histogram of z-scores in published papers. I'm always baffled by the discourse around it. Why is this plot a symptom of a problem? Under what model would we expect this to be a bell curve?
November 15, 2025 at 3:31 PM
Over on the Gelman blog, there's a post about this histogram of z-scores in published papers. I'm always baffled by the discourse around it. Why is this plot a symptom of a problem? Under what model would we expect this to be a bell curve?
Finished Gati’s “Zbig”. Got him in a Grand-Grand Strategy way. I’ve never experienced Bloodlands, I don’t have his intuition, just the fuzzzziest Anglo-Acadian intuition about the value sets that make institutions, reinforce, combine to cause power. Hinterlandish.
Good read. A few bangers.
Good read. A few bangers.
November 15, 2025 at 2:40 AM
Finished Gati’s “Zbig”. Got him in a Grand-Grand Strategy way. I’ve never experienced Bloodlands, I don’t have his intuition, just the fuzzzziest Anglo-Acadian intuition about the value sets that make institutions, reinforce, combine to cause power. Hinterlandish.
Good read. A few bangers.
Good read. A few bangers.
Reposted by Christopher Berry
@alphaxiv.org released quickarXiv
Swap arxiv → quickarxiv on any paper URL to get an instant blog with figures, insights, and explanations.
Swap arxiv → quickarxiv on any paper URL to get an instant blog with figures, insights, and explanations.
November 14, 2025 at 11:52 PM
@alphaxiv.org released quickarXiv
Swap arxiv → quickarxiv on any paper URL to get an instant blog with figures, insights, and explanations.
Swap arxiv → quickarxiv on any paper URL to get an instant blog with figures, insights, and explanations.
Reposted by Christopher Berry
According to signaling theory, some signals must be costly just to be costly—that's how you get a separating equilibrium. Think peacocks and their oversized feathers. So even if AI removes one costly signal, it doesn't mean we should stop technological progress — we'll just find new ones.
Is AI making job recruitment less meritocratic? We're getting some v interesting research studies on this question now, and the news is... not good. @jburnmurdoch.ft.com & I dive in, in the latest edition of our newsletter The AI Shift www.ft.com/content/e5b7...
November 14, 2025 at 11:52 AM
According to signaling theory, some signals must be costly just to be costly—that's how you get a separating equilibrium. Think peacocks and their oversized feathers. So even if AI removes one costly signal, it doesn't mean we should stop technological progress — we'll just find new ones.
The twist:
Colleague is a Turniptologist
Colleague is a Turniptologist
When I was at Arizona, a colleague teaching a class described this exchange:
Student: Where’s that idea in the textbook?
Colleague: This is from one of my own articles
S [angrily]: You mean you’re just MAKING IT UP?!
C: Do you think textbooks are dug up out of the ground like turnips or something?
Student: Where’s that idea in the textbook?
Colleague: This is from one of my own articles
S [angrily]: You mean you’re just MAKING IT UP?!
C: Do you think textbooks are dug up out of the ground like turnips or something?
November 14, 2025 at 1:05 AM
The twist:
Colleague is a Turniptologist
Colleague is a Turniptologist
“Moreover, we present a dataset comprising 25,368 samples (CausalDR), where each sample includes an input question, explicit causal DAG, graph-based reasoning trace, and validated answer.”
Researchers have developed CDCR-SFT, a novel method that enhances causal reasoning in language models, achieving an impressive 95.33% accuracy while significantly reducing logical hallucinations. https://arxiv.org/abs/2508.12495
Mitigating Hallucinations in Large Language Models via Causal Reasoning
ArXiv link for Mitigating Hallucinations in Large Language Models via Causal Reasoning
arxiv.org
November 14, 2025 at 1:02 AM
“Moreover, we present a dataset comprising 25,368 samples (CausalDR), where each sample includes an input question, explicit causal DAG, graph-based reasoning trace, and validated answer.”
“That said, eroding social cohesion, increasing polarization, and significant global events provide fertile ground for radicalization and many who turn to violence radicalize exclusively online…”
November 13, 2025 at 9:56 PM
“That said, eroding social cohesion, increasing polarization, and significant global events provide fertile ground for radicalization and many who turn to violence radicalize exclusively online…”
If you just woke up in ET, there’s 80 minutes, 12 flash talks coming up, then keynote, then the last sessions with industry. It’s better than your standing Thursday AM meetings.
causalscience.org
#CDSM25
causalscience.org
#CDSM25
Causal Data Science Meeting - Home
Fostering a dialogue between industry and academia on causal data science.
causalscience.org
November 13, 2025 at 1:26 PM
If you just woke up in ET, there’s 80 minutes, 12 flash talks coming up, then keynote, then the last sessions with industry. It’s better than your standing Thursday AM meetings.
causalscience.org
#CDSM25
causalscience.org
#CDSM25
Waking up to Marco Barbero Mot and Danilo Messinese’s preso on unsupervised discovery of causal mechanisms. Pleasantly surprised that they used clickstream data! #cdsm25
November 13, 2025 at 12:26 PM
Waking up to Marco Barbero Mot and Danilo Messinese’s preso on unsupervised discovery of causal mechanisms. Pleasantly surprised that they used clickstream data! #cdsm25
Day 2 of #CDSM25 kicks off at early ET / Centre Of The Universe Time. I’m aiming to be awake by “Unsupervised Discovery of Causal Mechanisms for Management Research” Marco Barbero Mot (Vanderbilt University) , Danilo Messinese (IE University)
www.causalscience.org
www.causalscience.org
Causal Data Science Meeting - Home
Fostering a dialogue between industry and academia on causal data science.
www.causalscience.org
November 12, 2025 at 9:48 PM
Day 2 of #CDSM25 kicks off at early ET / Centre Of The Universe Time. I’m aiming to be awake by “Unsupervised Discovery of Causal Mechanisms for Management Research” Marco Barbero Mot (Vanderbilt University) , Danilo Messinese (IE University)
www.causalscience.org
www.causalscience.org
Doing night three of The Hinton Lectures (TM). Likely with a lot of the usual #AIGS suspects.
After having been spoiled all day by the best intermission music in science (feat @p-hunermund.com at #CDSM25), I’m somewhat assaulted by the light Canadian hotel jazz for the pre-show.
After having been spoiled all day by the best intermission music in science (feat @p-hunermund.com at #CDSM25), I’m somewhat assaulted by the light Canadian hotel jazz for the pre-show.
November 12, 2025 at 9:43 PM
Doing night three of The Hinton Lectures (TM). Likely with a lot of the usual #AIGS suspects.
After having been spoiled all day by the best intermission music in science (feat @p-hunermund.com at #CDSM25), I’m somewhat assaulted by the light Canadian hotel jazz for the pre-show.
After having been spoiled all day by the best intermission music in science (feat @p-hunermund.com at #CDSM25), I’m somewhat assaulted by the light Canadian hotel jazz for the pre-show.
Reposted by Christopher Berry
An analysis of 47,000 publicly shared ChatGPT conversations: ~10% related to emotional or mental health, ChatGPT exhibits a "default to yes" behavior, and more (Washington Post)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
November 12, 2025 at 12:36 PM
An analysis of 47,000 publicly shared ChatGPT conversations: ~10% related to emotional or mental health, ChatGPT exhibits a "default to yes" behavior, and more (Washington Post)
Main Link | Techmeme Permalink
Main Link | Techmeme Permalink
“Based on spatial distribution of ATP and explored areas, Physarum can perform up to ~10^36 logical operations in 24 hours, scaling linearly in the non-equilibrium steady state.“
New preprint:
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
Morphological computational capacity of Physarum polycephalum
While computational capacity limits of the universe and carbon-based life have been estimated, a stricter bound for aneural organisms has not been established. Physarum polycephalum, a unicellular, mu...
arxiv.org
November 12, 2025 at 2:39 AM
“Based on spatial distribution of ATP and explored areas, Physarum can perform up to ~10^36 logical operations in 24 hours, scaling linearly in the non-equilibrium steady state.“
Reposted by Christopher Berry
New preprint:
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
Morphological computational capacity of Physarum polycephalum
While computational capacity limits of the universe and carbon-based life have been estimated, a stricter bound for aneural organisms has not been established. Physarum polycephalum, a unicellular, mu...
arxiv.org
November 10, 2025 at 2:16 PM
New preprint:
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
arxiv.org/abs/2510.19976
"Morphological computational capacity of Physarum polycephalum"
Suyash Bajpai, Aviva Lucas-DeMott, @msahsorin.bsky.social, Philip Kurian
What a read Tanya Talaga’s “The Knowing” is. There’s a passage about identity and persistence that hit.
I don’t know how I’d recommend the sequence. Probably do The North-West Is Our Mother first, as it’ll make the Métis references in Northern Ontario gel a bit maybe.
I don’t know how I’d recommend the sequence. Probably do The North-West Is Our Mother first, as it’ll make the Métis references in Northern Ontario gel a bit maybe.
November 12, 2025 at 12:02 AM
What a read Tanya Talaga’s “The Knowing” is. There’s a passage about identity and persistence that hit.
I don’t know how I’d recommend the sequence. Probably do The North-West Is Our Mother first, as it’ll make the Métis references in Northern Ontario gel a bit maybe.
I don’t know how I’d recommend the sequence. Probably do The North-West Is Our Mother first, as it’ll make the Métis references in Northern Ontario gel a bit maybe.
Down at the centre of the Universe for The Hinton Lectures.
November 11, 2025 at 9:44 PM
Down at the centre of the Universe for The Hinton Lectures.
Reposted by Christopher Berry
To give you an update, I have to do a little background on decoding. 🧵
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
November 10, 2025 at 6:57 PM
To give you an update, I have to do a little background on decoding. 🧵
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
At the end of the forward pass, the model produces a last hidden state, which is a vector. Cosine similarity measures how parallel that vector is to all the token embeddings. Each token gets a score called a logit based on this.
Back at Shopify tonight. 10th floor. AI Tinkerers is good one, grassroots, quite a few serious builders come out.
November 10, 2025 at 10:44 PM
Back at Shopify tonight. 10th floor. AI Tinkerers is good one, grassroots, quite a few serious builders come out.
I support both public broadcasting and a system of private competition. Both the public system and private system (of tainted access journalism) has its own pros and cons. To be consistently pluralist, both need to thrive for us all to flourish.
I reckon a shared national popular culture is just as important (if not more important) than a fact based, impartial news service for liberal democracy to work.
November 10, 2025 at 3:26 PM
I support both public broadcasting and a system of private competition. Both the public system and private system (of tainted access journalism) has its own pros and cons. To be consistently pluralist, both need to thrive for us all to flourish.
Packed floor in spite of the snow today. Demo was well received in that the signal hit the intended target and I got a lot of real feedback. Appreciative for those gifts. And the disclosure of thoughtcrime.
Friends, you know who you are, thank you!
Friends, you know who you are, thank you!
November 10, 2025 at 3:33 AM
Packed floor in spite of the snow today. Demo was well received in that the signal hit the intended target and I got a lot of real feedback. Appreciative for those gifts. And the disclosure of thoughtcrime.
Friends, you know who you are, thank you!
Friends, you know who you are, thank you!
Looking forward to this one in particular: “Forests for Differences: Robust Causal Inference Beyond Parametric DiDHugo Gobato Souto (Luizalabs) and Francisco Louzada Neto (University of Sao Paolo)”
#CDSM2025
#CDSM2025
November 10, 2025 at 2:16 AM
Looking forward to this one in particular: “Forests for Differences: Robust Causal Inference Beyond Parametric DiDHugo Gobato Souto (Luizalabs) and Francisco Louzada Neto (University of Sao Paolo)”
#CDSM2025
#CDSM2025
Maybe there are 8 aubergine democratic senators? And maybe they’re capable of calculus. Maybe some of them are blundering by misestimating the cleavage and the valence of the situation?
youtu.be/qHhsc4lYamc?...
youtu.be/qHhsc4lYamc?...
Eaton's Aubergine Commercial
YouTube video by littlepiaf
youtu.be
November 10, 2025 at 2:10 AM
Maybe there are 8 aubergine democratic senators? And maybe they’re capable of calculus. Maybe some of them are blundering by misestimating the cleavage and the valence of the situation?
youtu.be/qHhsc4lYamc?...
youtu.be/qHhsc4lYamc?...